RCAN-DETR: A PCB Defect Detection Model Based on Re-calibrated Multi-scale Fusion

Abstract To tackle the precision bottleneck in Printed Circuit Board (PCB) defect detection, which arises from highly similar target textures and minuscule defect sizes, this study presents RCAN-DETR, an enhanced model based on RT-DETR. Firstly, a multi-branch re-parameterization module, Conv3XC, is designed to efficiently extract edge and structural information from images via parallel branches, thereby strengthening the capability to capture semantic features.Secondly, an innovative Re-Calibration Feature Pyramid Network (Re-Calibration FPN) is developed. This network aggregates fine-grained boundary details and rich semantic information to generate accurate object contour representations, while dynamically recalibrating predictions of object locations.Thirdly, a Deformable Multi-Head Attention (DMHA) module is proposed, which leverages its dynamic focusing capability to efficiently capture complex relationships within the feature space.Finally, the NWD (Normalized Wasserstein Distance) and GIoU (Generalized Intersection over Union) loss functions are integrated to optimize the recognition and matching of small targets, while accelerating the convergence rate.Experimental evaluations on the publicly available Peking University PCB defect dataset demonstrate that, in comparison with the baseline model, the proposed model achieves improvements of 3.2%, 4.0%, 5.5%, and 2.3% in Precision (P), Recall (R), mean Average Precision (mAP), and mAP@0.5:0.95, respectively. The enhanced algorithm markedly improves PCB defect detection performance, offering an innovative solution for deploying high-precision detection systems.